Re: Mangling problem?

[Greg B. <gn...@sg...>]

On Fri, Oct 15, 2004 at 07:19:42PM +0100, John Levon wrote:
> On Sat, Oct 16, 2004 at 04:01:19AM +1000, Greg Banks wrote:
> 
> > Ok, so that was a wild goose chase.  The real problem I'm trying
> > to track down is that my backtrace samples are not being emitted to
> > the gmon.out file by opgprof.
> 
> Can I see your backtrace kernel patch ?

Sure, attached (actually a slightly newer one than I'm running).  I
was going to post it when I was sure it was working with a modern
kernel and oprofile 0.8.1.  It has previously worked with kernel
2.6.4 and oprofile-cvs20040504.  Tracing work indicates it's
producing correct sequences of PC values.


> > KERNEL_ENTER_SWITCH to kernel
> > Lookup of cookie 0 failed, errno=22
> > CTX_SWITCH to tid 0, tgid 0, cookie 0(none)
> > Opening "/var/lib/oprofile/samples/current/{root}//root/vmlinux-2.6.9-rc3-gnbm/{dep}/{root}//root/vmlinux-2.6.9-rc3-gnbm/CPU_CYCLES.1000000.0.all.all.all"
> > Sample at 0x1a960(0): invalid(ffffffffffffffff), app invalid(ffffffffffffffff), kernel /root/vmlinux-2.6.9-rc3-gnbm
> > Opening "/var/lib/oprofile/samples/current/{root}//root/vmlinux-2.6.9-rc3-gnbm/{dep}/{root}//root/vmlinux-2.6.9-rc3-gnbm/{cg}/{root}//root/vmlinux-2.6.9-rc3-gnbm/CPU_CYCLES.1000000.0.all.all.all"
> > Arc at 0x1a960(0): invalid(ffffffffffffffff), app invalid(ffffffffffffffff), kernel /root/vmlinux-2.6.9-rc3-gnbm
> > Arc at 0x1ac60(0): invalid(ffffffffffffffff), app invalid(ffffffffffffffff), kernel /root/vmlinux-2.6.9-rc3-gnbm
> 
> This all looks pretty broken.
> 
> > callgraph filenames:
> > profile_dep_set #1:
> > lib_image: /root/vmlinux-2.6.9-rc3-gnbm
> > profile_sample_files #0:
> > sample_filename: 
> > callgraph filenames:
> > /var/lib/oprofile/samples/current/{root}/root/vmlinux-2.6.9-rc3-gnbm/{dep}/{root}/root/vmlinux-2.6.9-rc3-gnbm/{cg}/{root}/root/vmlinux-2.6.9-rc3-gnbm/CPU_CYCLES.1000000.0.all.all.all
> 
> Looks OK too. I'll have to check later what's expected. I didn't really
> understand your comments above

There are two profile_dep_sets, #0 has a flat sample filename and no
callgraph sample filename, and #1 has vice versa.  Indented:

profile_classes:
    event: 
    cpuinfo: CPU: Itanium 2, speed 1300 MHz (estimated)
    class #0:
	name: CPU_CYCLES:1000000
	longname: Counted CPU_CYCLES events (CPU Cycles) with a unit mask of 0x00 (No unit mask) count 1000000
	ptemplate:
	    event: CPU_CYCLES
	    count: 1000000
	    unitmask: 
	    tgid: 
	    tid: 
	    cpu: 
	profiles_set #0:
	    image: /root/vmlinux-2.6.9-rc3-gnbm
	    profile_dep_set #0:
		lib_image: /root/vmlinux-2.6.9-rc3-gnbm
		profile_sample_files #0:
		    sample_filename: /var/lib/oprofile/samples/current/{root}/root/vmlinux-2.6.9-rc3-gnbm/{dep}/{root}/root/vmlinux-2.6.9-rc3-gnbm/CPU_CYCLES.1000000.0.all.all.all
		    callgraph filenames:
	    profile_dep_set #1:
		lib_image: /root/vmlinux-2.6.9-rc3-gnbm
		profile_sample_files #0:
		    sample_filename: 
		    callgraph filenames:
			/var/lib/oprofile/samples/current/{root}/root/vmlinux-2.6.9-rc3-gnbm/{dep}/{root}/root/vmlinux-2.6.9-rc3-gnbm/{cg}/{root}/root/vmlinux-2.6.9-rc3-gnbm/CPU_CYCLES.1000000.0.all.all.all

Is that right?  It looks suspicious to me, I was expecting a
single profile_dep_set, like this:

profile_classes:
    event: 
    cpuinfo: CPU: Itanium 2, speed 1300 MHz (estimated)
    class #0:
	name: CPU_CYCLES:1000000
	longname: Counted CPU_CYCLES events (CPU Cycles) with a unit mask of 0x00 (No unit mask) count 1000000
	ptemplate:
	    event: CPU_CYCLES
	    count: 1000000
	    unitmask: 
	    tgid: 
	    tid: 
	    cpu: 
	profiles_set #0:
	    image: /root/vmlinux-2.6.9-rc3-gnbm
	    profile_dep_set #0:
		lib_image: /root/vmlinux-2.6.9-rc3-gnbm
		profile_sample_files #0:
		    sample_filename: /var/lib/oprofile/samples/current/{root}/root/vmlinux-2.6.9-rc3-gnbm/{dep}/{root}/root/vmlinux-2.6.9-rc3-gnbm/CPU_CYCLES.1000000.0.all.all.all
		    callgraph filenames:
			/var/lib/oprofile/samples/current/{root}/root/vmlinux-2.6.9-rc3-gnbm/{dep}/{root}/root/vmlinux-2.6.9-rc3-gnbm/{cg}/{root}/root/vmlinux-2.6.9-rc3-gnbm/CPU_CYCLES.1000000.0.all.all.all


Greg.
-- 
Greg Banks, R&D Software Engineer, SGI Australian Software Group.
I don't speak for SGI.